Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Decoupling Bidirectional Geometric Representations of 4D cost volume with 2D convolution

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Pure 2D convolutions match 3D cost-volume accuracy in real-time stereo.

desk verdict A plausible real-time stereo architecture whose central decoupling claim hinges on an unstated grouping detail; worth review but not citable as-is. read the letter →

arxiv 2509.02415 v1 pith:CZCMKSOZ submitted 2025-09-02 cs.CV

Xiaobao Wei , Changyong Shu , Zhaokun Yue , Chang Huang , Weiwei Liu , Shuai Yang , Lirong Yang , Peng Gao
show 3 more authors
Wenbin Zhang Gaochao Zhu Chengxiang Wang
This is my paper
classification cs.CV
keywords stereomatching4Dcostvolume2Dconvolutionaggregationdisparityestimationreal-timeinferencedecoupledgroup-wisecorrelation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the standard 3D-convolution regularization step in stereo matching can be replaced by two cheap 2D convolution steps without losing accuracy. It reshapes the 4D cost volume (disparity × channels × height × width) into a 3D tensor, then separates aggregation into a spatial-smoothness step and a disparity-unimodality step, each a pure 2D operation. The resulting network, DBStereo, reports 0.45 px end-point error on Scene Flow at 49 ms, beating the iterative method IGEV-Stereo (0.47 px, 340 ms) and all listed aggregation-based methods on both accuracy and speed. If true, high-accuracy stereo matching no longer needs 3D convolutions, which would make real-time edge and mobile deployment far more practical.

What carries the argument

The load-bearing mechanism is the Channel2Disp reshape: the 4D group-wise correlation volume is flattened from (D × G × H × W) to (G·D) × H × W, making the disparity dimension a channel dimension. Two modules then process this tensor with pure 2D convolutions: a Spatial Aggregation module (a 2D convolution, e.g., 3×3, over H×W) that realizes the local-smoothness prior, and a Disparity Aggregation module (a 2D 1×1 convolution) that performs a global, fully-connected-style interaction across the G·D channels at each pixel, realizing the disparity-unimodality prior. These modules are stacked inside the Bidirectional Geometry Aggregation (BGA) block with an encoder-decoder structure. The argumen

What would settle it

Run the BGA block with a strictly depthwise or grouped 3×3 convolution (groups = G·D) so spatial aggregation cannot mix disparity channels, then compare Scene Flow EPE and D1 against the reported DBStereo numbers. If accuracy drops sharply, the gains come from cross-disparity channel mixing rather than from the claimed decoupled priors; if accuracy holds, the decoupling is real.

Watch

Extended reading notes

Core claim

The paper's central claim is that the 4D cost volume in stereo matching, traditionally regularized by stacks of 3D convolutions, can be reshaped into a 3D tensor and aggregated by two successive 2D convolution steps without losing accuracy. The first step, spatial aggregation, applies a 2D convolution over the image plane to enforce spatial local smoothness; the second, disparity aggregation, applies a 1×1 convolution that mixes all disparity channels at each pixel, enforcing a sharp unimodal disparity distribution. The authors argue that this decoupling encodes two task-specific priors directly into the architecture, removing the need for coupled, slow-receptive-field 3D convolutions. On Sc

Load-bearing premise

The whole decoupling story rests on the 3×3 2D convolution acting as a purely spatial aggregation that does not mix disparity information, but the paper never specifies whether that convolution is grouped or depthwise; if it is a standard cross-channel convolution, the separation between spatial and disparity aggregation is approximate rather than real.

Editorial extensions

If this is right

  • High-accuracy stereo matching can run on edge and mobile hardware: DBStereo-S runs at 15 ms with 0.65 EPE, and DBStereo-L runs at 49 ms with 0.45 EPE on Scene Flow.
  • The field's assumption that 4D cost volumes must be aggregated by dimension-matched 3D convolutions is broken, opening a research direction for pure-2D cost aggregation.
  • The explicit spatial-smoothness and disparity-unimodality inductive biases let a lightweight model match iterative-refinement accuracy with a single-stage smooth-L1 training loss.
  • Aggregation-based methods can outperform iterative optimization methods on both speed and accuracy, reversing the usual trade-off.
  • DBStereo provides a simple baseline that future decoupled-aggregation designs can build on and compare against.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: The decoupling principle likely transfers beyond stereo: any matching or estimation task with a 4D cost volume and a global-unimodality prior along one axis (optical flow range, multi-view depth, video disparity) could use the same reshape-and-2D trick.
  • Inference: Since the paper never specifies whether the spatial 3×3 convolutions are grouped or depthwise, an immediate test is to force strict decoupling with groups = G·D; if accuracy collapses, the reported gains come from cross-disparity channel mixing rather than from the stated spatial-disparity separation.
  • Inference: The 1×1 disparity aggregation is effectively a per-pixel fully connected layer over the disparity dimension, which connects this design to attention or MLP-mixer architectures and suggests replacing it with a learned global disparity transform to sharpen unimodality further.
  • Inference: A theoretical account of why G·D channels suffice to encode 4D geometry is missing; proving an equivalence or approximation bound between 3D convolutions and this factored 2D form would solidify the new paradigm.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DBStereo, a stereo matching network that replaces 3D cost-volume regularization with pure 2D convolutions. The 4D group-wise cost volume is reshaped to a (G·D)×H×W tensor, and a Bidirectional Geometry Aggregation (BGA) block stacks 2D spatial-aggregation convolutions (3×3) with 1×1 disparity-aggregation convolutions, motivated by spatial-smoothness and disparity-unimodality priors. On Scene Flow (Table 1), the largest variant DBStereo-L reports EPE 0.45, D1 1.57%, and 49 ms runtime, outperforming the compared aggregation-based methods and the iterative IGEV-Stereo. The paper claims state-of-the-art accuracy and real-time speed on multiple benchmarks, but KITTI results are mentioned in the text and never shown, no ablations isolate the two aggregation modules, and the grouping of the spatial 3×3 convolution is not specified. The central architectural claim therefore remains under-specified as written.

Significance. If the decoupled aggregation is implemented as claimed, the paper would provide a meaningful challenge to the assumption that 4D cost volumes require 3D convolutions, and it would make high-accuracy real-time stereo more plausible on edge devices. The Scene Flow numbers in Table 1 are competitive: DBStereo-L achieves lower EPE and D1 than IGEV-Stereo at roughly one-seventh the runtime. The task-specific priors are clearly stated and are plausible inductive biases. However, the paper does not yet establish the core claim because the architecture description is ambiguous on the crucial grouping of the spatial convolution, and because the reported evidence is limited to Scene Flow without ablations or KITTI results. The manuscript promises code but does not provide it, which further hinders verification.

major comments (4)
  1. [§3.2, §4.3 (Eq. 2)] The central decoupling claim depends on the 3×3 spatial-aggregation convolution acting 'within the same disparity level.' This is only true if the convolution is grouped so that channels from different disparity hypotheses do not interact. The manuscript never specifies groups. With the default Conv2d groups=1, each output channel is a sum over all G·D input channels, so the 3×3 kernel already mixes disparity candidates, and the subsequent 1×1 convolution is not the first point of disparity mixing. The architecture would then collapse to an ordinary 2D CNN on a reshaped volume, and the inductive-bias argument in §3 is unsupported. Please specify the grouping (e.g., groups=D or depthwise), state the kernel connectivity explicitly, and update Figure 2 accordingly. If no grouping is used, the decoupling claim must be revised.
  2. [§5.3] The text states that results are submitted to the KITTI 2012 and KITTI 2015 leaderboards and that 'EPE and >3px metric' are reported for cross-domain generalization, but no KITTI table or numerical results appear anywhere in the manuscript. Since the abstract and contributions claim state-of-the-art performance on 'multiple benchmarks' and 'outperforms all existing aggregation-based methods,' the missing KITTI evidence is load-bearing. Please include the KITTI 2012/2015 results and cross-domain generalization numbers, or narrow the claim to the Scene Flow benchmark actually reported.
  3. [§4.3, Table 1] No ablation isolates the contribution of the Spatial Aggregation module from the Disparity Aggregation module, nor does any ablation vary the BGA stacking, channel widths, or the Channel2Disp reshape. Table 1 only compares three full-network variants (S/M/L), so the attribution of the accuracy gains to the proposed decoupled inductive biases is not established. Add ablations that remove or reorder the two modules, and compare against a 3D-convolution baseline using the same MobileNetV2 backbone and training schedule.
  4. [§4.1–4.4] Several components needed for reproducibility are unspecified: the upsampling blocks after the pretrained backbone, the generation and insertion of spatial attention from Fl,4/Fl,8/Fl,16, the encoder–decoder structure of the BGA block, the number of stacked spatial/disparity modules, and the 'learnable parameters' used for disparity upsampling. Since code is promised but not yet available, these omissions cannot be resolved from the text. Please provide a detailed architecture table or pseudocode.
minor comments (5)
  1. [References] Duplicate or near-duplicate entries appear: [5] and [6] are both DeepPruner; [15] and [16] are both RAFT-Stereo; [30] and [31] are both IGEV-Stereo. Please consolidate.
  2. [Notation] Section 3.2 uses (D·C×H×W) for the reshaped volume while Eq. (2) uses (G·D)×H×W. Clarify whether C refers to the number of feature channels or the number of groups in the group-wise correlation volume, and use one notation consistently.
  3. [Abstract / Figure 1] The phrase 'outperforms all existing aggregation-based methods' is too broad for the experiments shown. Table 1 compares a subset of methods; please qualify the claim as 'the compared aggregation-based methods' or provide a more exhaustive comparison.
  4. [Table 1] The runtime for BANET-2D is listed as 'xx', which is incomplete. Also, no parameter count or FLOPs are reported for DBStereo-S/M/L, making the real-time/deployment claims difficult to assess.
  5. [Throughout] There are numerous typos and grammatical errors, e.g., 'mathcing', 'Sterero', 'break the empirical design', 'utlized', 'concatenat', 'inferenc time', and 'D1 outlier with an absolute error greater than 1 pixels.' A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture is motivated by priors and validated on external benchmarks; the sole self-citation is not load-bearing.

full rationale

The paper's central claim is empirical: that a pure-2D decoupled aggregation block reaches state-of-the-art accuracy on SceneFlow and KITTI. The derivation in Section 3.2 starts from two explicitly stated priors (spatial local smoothness and disparity unimodality) and proposes a reshape plus two 2D convolution steps as an architectural embodiment of those priors. This is an inductive-bias design choice, not a derivation of a quantity from itself. Equation (2) is a reshape; Eq. (3) is the standard softmax disparity regression; Eq. (4) is a standard smooth-L1 loss. None of these equations reduces to a fitted parameter that is then renamed as a prediction. The only self-citation is reference [25], listed among iterative optimization paradigms in the introduction; it is not used to justify the BGA block, the decoupling claim, or any benchmark result, so it is not load-bearing. A potential concern that the 3x3 spatial convolution might mix disparity channels if not grouped is an implementation/correctness ambiguity, not circularity, because the reported numbers come from external benchmark evaluation rather than from the priors themselves. Therefore the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no physical entities. Its central claim rests on two task priors, a mobile backbone, and undisclosed architecture hyperparameters. The most consequential unstated choice, the grouping of the 2D convolutions, determines whether the claimed spatial/disparity decoupling is actually realized.

free parameters (4)
  • Loss weights lambda0 and lambda1 = lambda0=0.3, lambda1=1
    Chosen by hand in Eq 4; they affect the final disparity quality and are not derived.
  • Maximum disparity Dmax = not stated
    Appears in Eq 3 as Dmax/4; its value determines cost volume size and runtime but is never reported.
  • BGA architecture hyperparameters = not stated
    Number of spatial and disparity aggregation modules, channels, groups, and kernel groupings are not disclosed; these directly determine the reported accuracy and speed.
  • Training epochs = 90 SceneFlow, 500 KITTI fine-tune
    Training schedule is chosen by hand and contributes to the reported performance.
assumptions (5)
  • domain assumption Spatial Local Smoothness Prior: adjacent pixels at the same depth have similar disparity values.
    Invoked in Sec 3.2 to justify the spatial 2D aggregation module.
  • domain assumption Disparity Unimodality Prior: for a single pixel the disparity probability distribution should be a sharp unimodal distribution.
    Invoked in Sec 3.2 to justify the disparity 1x1 aggregation module.
  • standard math A 1x1 2D convolution over the reshaped (G·D) channel dimension provides global disparity aggregation.
    True as a linear mixing across D at each spatial location, but only if the preceding spatial convolutions are grouped so they do not mix disparity channels; the grouping is never specified.
  • domain assumption MobileNetV2 pretrained on ImageNet transfers useful features for stereo matching.
    Used in Sec 4.1 as the feature extractor without additional evidence.
  • domain assumption SceneFlow training and KITTI fine-tuning transfer to the leaderboard benchmarks without architectural changes.
    Implied by the evaluation protocol in Sec 5.2 and Sec 5.3, but no KITTI numbers are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupling Bidirectional Geometric Representations of 4D cost volume with 2D convolution." pith.science (2026). https://pith.science/paper/CZCMKSOZ

@misc{pith2026250902415,
  author       = {Pith},
  title        = {Pith review of: Decoupling Bidirectional Geometric Representations of 4D cost volume with 2D convolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZCMKSOZ}},
  note         = {Machine review of arXiv:2509.02415}
}
read the original abstract

High-performance real-time stereo matching methods invariably rely on 3D regularization of the cost volume, which is unfriendly to mobile devices. And 2D regularization based methods struggle in ill-posed regions. In this paper, we present a deployment-friendly 4D cost aggregation network DBStereo, which is based on pure 2D convolutions. Specifically, we first provide a thorough analysis of the decoupling characteristics of 4D cost volume. And design a lightweight bidirectional geometry aggregation block to capture spatial and disparity representation respectively. Through decoupled learning, our approach achieves real-time performance and impressive accuracy simultaneously. Extensive experiments demonstrate that our proposed DBStereo outperforms all existing aggregation-based methods in both inference time and accuracy, even surpassing the iterative-based method IGEV-Stereo. Our study break the empirical design of using 3D convolutions for 4D cost volume and provides a simple yet strong baseline of the proposed decouple aggregation paradigm for further study. Code will be available at (\href{https://github.com/happydummy/DBStereo}{https://github.com/happydummy/DBStereo}) soon.

Figures

Figures reproduced from arXiv: 2509.02415 by the authors.

Figure 1
Figure 1. The proposed DBStereo decouple the traditional 3D aggregation into spatial aggregation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of our proposed DBStereo. The Bidirectional Geometry Aggregation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multi-Resolution Voxelized Map-Based Stereo Visual-Inertial Odometry

    cs.RO 2026-07 conditional novelty 6.0 of 10

    MR-Voxel-SVIO combines multi-resolution voxelized LiDAR maps with depth-adaptive 3D-DDA ray casting to achieve accurate edge-cloud stereo VIO with very low data transfer.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bangunharcana, J

    A. Bangunharcana, J. W. Cho, S. Lee, I. S. Kweon, K.-S. Kim, and S. Kim. Correlate-and- excite: Real-time stereo matching via guided cost volume excitation. In 2021 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS) , pages 3542–3548. IEEE, 2021

    work page 2021

  2. [2]

    Chang and Y .-S

    J.-R. Chang and Y .-S. Chen. Pyramid stereo matching network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    work page 2018

  3. [3]

    Cheng, W

    J. Cheng, W. Yin, K. Wang, X. Chen, S. Wang, and X. Yang. Adaptive fusion of single-view and multi-view depth for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10138–10147, 2024. 7

    work page 2024

  4. [4]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchi- cal image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition , pages 248–255, 2009

    work page 2009

  5. [5]

    Duggal, S

    S. Duggal, S. Wang, W.-C. Ma, R. Hu, and R. Urtasun. Deeppruner: Learning efficient stereo matching via differentiable patchmatch. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 4384–4393, 2019

    work page 2019

  6. [6]

    Duggal, S

    S. Duggal, S. Wang, W.-C. Ma, R. Hu, and R. Urtasun. Deeppruner: Learning efficient stereo matching via differentiable patchmatch. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4384–4393, 2019

    work page 2019

  7. [7]

    Duggal, S

    S. Duggal, S. Wang, W.-C. Ma, R. Hu, and R. Urtasun. Deeppruner: Learning efficient stereo matching via differentiable patchmatch. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), October 2019

    work page 2019

  8. [8]

    Geiger, P

    A. Geiger, P. Lenz, and R. Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition , pages 3354–3361. IEEE, 2012

    work page 2012

Show all 37 references
  1. [9]

    X. Guo, K. Yang, W. Yang, X. Wang, and H. Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  2. [10]

    Hsieh and S.-S

    Y .-Z. Hsieh and S.-S. Lin. Robotic arm assistance system based on simple stereo matching and q-learning optimization. IEEE Sensors Journal, 20(18):10945–10954, 2020

    2020

  3. [11]

    Kendall, H

    A. Kendall, H. Martirosyan, S. Dasgupta, P. Henry, R. Kennedy, A. Bachrach, and A. Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017

    2017

  4. [12]

    Khamis, S

    S. Khamis, S. Fanello, C. Rhemann, A. Kowdle, J. Valentin, and S. Izadi. Stereonet: Guided hierarchical refinement for real-time edge-aware depth prediction. In Proceedings of the Euro- pean conference on computer vision (ECCV), pages 573–590, 2018

    2018

  5. [13]

    X. Li, C. Zhang, W. Su, and W. Tao. Iinet: Implicit intra-inter information fusion for real-time stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3225–3233, 2024

    2024

  6. [14]

    Liang, Y

    Z. Liang, Y . Guo, Y . Feng, W. Chen, L. Qiao, L. Zhou, J. Zhang, and H. Liu. Stereo matching using multi-level cost volume and multi-scale feature constancy.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):300–315, 2019

    2019

  7. [15]

    Lipson, Z

    L. Lipson, Z. Teed, and J. Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218–227, 2021

    2021

  8. [16]

    Lipson, Z

    L. Lipson, Z. Teed, and J. Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218–227. IEEE, 2021

    2021

  9. [17]

    Mayer, E

    N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 4...

    2016

  10. [18]

    Menze and A

    M. Menze and A. Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3061–3070, 2015

    2015

  11. [19]

    Nie, M.-M

    G.-Y . Nie, M.-M. Cheng, Y . Liu, Z. Liang, D.-P. Fan, Y . Liu, and Y . Wang. Multi-level con- text ultra-aggregation for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3283–3291, 2019

    2019

  12. [20]

    Sandler, A

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  13. [21]

    Shamsafar, S

    F. Shamsafar, S. Woerz, R. Rahim, and A. Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision (WACV), pages 2417–2426, January 2022

    2022

  14. [22]

    Tankovich, C

    V . Tankovich, C. Hane, Y . Zhang, A. Kowdle, S. Fanello, and S. Bouaziz. Hitnet: Hierar- chical iterative tile refinement network for real-time stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14362–14372, 2021. 8

    2021

  15. [23]

    Q. Wang, S. Shi, S. Zheng, K. Zhao, and X. Chu. Fadnet++: Real-time and accurate disparity estimation with configurable networks. arXiv preprint arXiv:2110.02582, 2021

    2021 arXiv

  16. [24]

    X. Wang, G. Xu, H. Jia, and X. Yang. Selective-stereo: Adaptive frequency information se- lection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701–19710, 2024

    2024

  17. [25]

    X. Wei, J. Liu, D. Yang, J. Cheng, C. Shu, and W. Wang. A wavelet-based stereo matching framework for solving frequency convergence inconsistency. arXiv preprint arXiv:2505.18024, 2025

    2025 arXiv

  18. [26]

    B. Wen, M. Trepte, J. Aribido, J. Kautz, O. Gallo, and S. Birchfield. Foundationstereo: Zero- shot stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5249–5260, June 2025

    2025

  19. [27]

    Z. Wu, X. Wu, X. Zhang, S. Wang, and L. Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 7484–7493, 2019

    2019

  20. [28]

    G. Xu, J. Cheng, P. Guo, and X. Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12981–12990, June 2022

    2022

  21. [29]

    G. Xu, J. Liu, X. Wang, J. Cheng, Y . Deng, J. Zang, Y . Chen, and X. Yang. Banet: Bilateral aggregation network for mobile stereo matching. arXiv preprint arXiv:2503.03259, 2025

    2025 arXiv

  22. [30]

    G. Xu, X. Wang, X. Ding, and X. Yang. Iterative geometry encoding volume for stereo match- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21919–21928, June 2023

    2023

  23. [31]

    G. Xu, X. Wang, X. Ding, and X. Yang. Iterative geometry encoding volume for stereo match- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919–21928, 2023

    2023

  24. [32]

    G. Xu, Y . Wang, J. Cheng, J. Tang, and X. Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 46(4):2461–2474, 2023

    2023

  25. [33]

    Xu and J

    H. Xu and J. Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1959–1968, 2020

    1959

  26. [34]

    Xu and J

    H. Xu and J. Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  27. [35]

    G. Yang, X. Song, C. Huang, Z. Deng, J. Shi, and B. Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899–908, 2019

    2019

  28. [36]

    Zenati and N

    N. Zenati and N. Zerhouni. Dense stereo matching with application to augmented reality. In 2007 IEEE International Conference on Signal Processing and Communications, pages 1503–

    2007

  29. [37]

    Zheng, X.-M

    D. Zheng, X.-M. Wu, Z. Liu, J. Meng, and W.-s. Zheng. Diffuvolume: Diffusion model for volume based stereo matching. International Journal of Computer Vision, 133(7):3807–3821, 2025. 9

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.